Researchers have developed a method to detect and remove backdoors in diffusion models used for AI-generated content. Diffusion models are vulnerable to attacks that insert malicious code into the training data, which can be exploited by hackers. The new approach, called Temporal Noise Consistency (TNC), identifies anomalies in the model's behavior when it is exposed to a backdoor trigger. TNC uses a combination of detection and detoxification techniques to remove the backdoor without compromising the model's performance. The method has been tested on five different types of backdoor attacks and outperforms existing defenses in terms of accuracy and efficiency. --- Why it matters: This matters because diffusion models are widely used for AI-generated content, such as images and videos, but they can be vulnerable to backdoor attacks that compromise their security. This research provides a new tool for detecting and removing these types of attacks, which is essential for maintaining the trustworthiness of AI systems. Source: https://arxiv.org/abs/2602.01765